189 ON THE VOLUME OF THE DOUBLE STOCHASTIC MATRICES 1. Introduction and Notation

advertisement
189
Acta Math. Univ. Comenianae
Vol. LXI, 2(1992), pp. 189–192
ON THE VOLUME OF THE
DOUBLE STOCHASTIC MATRICES
M. SCHMUCKENSCHLÄGER
1. Introduction and Notation
Let
Q
n
be the group of permutation matrices in Rn . Then
Dn : = co Π : Π ∈
!
Y
n
2
is a convex set in Rn . It is well known that Dn is the set of all double stochastic
matrices, i.e.




X
X
Dn = T = (tij )ni,j=1 :
tij =
tij = 1 ∀ i, j ∈ {1, . . . , n}, tij ∈ [0, 1]


i=1
j=1
The volume of Dn is somehow related to a Kahane type inequality (cf. [S]) for the
group of permutations, more precisely: let (xj,k ) be a double sequence in some
Banach space X, if
1
√
Volk (Dn ) k
≥c n
2
Volk (Bk )
2
where k is the dimension of Dn ⊆ Rn , then the L1 -norm and the norm associated
P
with ψ1 (t): = et −1 of || j,k xj,k πj,k || (the expectation being taken with respect to
the normalized counting measure on the group of all signed permutation matrices
(πj,k ) i.e. πj,k ∈ {−1, 0, 1}) are equivalent. Conversely, if the L1 -norm and the
2
norm associated with ψ2 (t): = et − 1 are equivalent, then the volume of Dn must
satiesfy the above inequality up to some logarithmic factor. We prove that such
an inequality can not hold. We also include a proof of an upper estimate for the
volume of a convex polytope all of whose vertices are at a given distance from the
origin. Though this result is known we could not find a reference.
Received December 11, 1991; revised October 1, 1992.
1980 Mathematics Subject Classification (1991 Revision). Primary 15A51; Secondary 52A22.
190
M. SCHMUCKENSCHLÄGER
2
It is easy to see that the subspace E of Rn defined by
E=
2n n
o
\
2
x ∈ Rn : hx, Nj i = 0
j=1
where
N1 = (1, . . . , 1, 0, . . . , 0),
| {z }
n
N2 = (0, . . . , 0, 1, . . . , 1, 0, . . . , 0), . . .
| {z } | {z }
n
n
Nn = (0, . . . , 0, 1, . . . , 1),
| {z }
n
Nn+1 = (1, 0, . . . , 0, 1, 0, . . . , 0, 1, . . . ),
| {z } | {z }
n−1
n−1
Nn+2 = (0, 1, 0, . . . , 0, 1, 0, . . . , 0, 1, . . . ), . . .
| {z } | {z }
n−1
n−1
N2n = (0, . . . , 0, 1, 0, . . . , 0, 1, . . . )
| {z } | {z }
n−1
n−1
has dimension (n − 1)2 . Thus the dimension of Dn is (n − 1)2 .
2. The Basic Estimates
In order to estimate the (n−1)2 -dimensional volume of Dn we need some results.
The first one is due to Vaaler [V] (a generalization of this result can be found in
[M-P]).
Lemma 2.1. Let E be a k-dimensional subspace of Rn . Then
Volk (Bn∞ ∩ E) ≥ 2k
where Bn∞ is the cube [−1, 1]n.
The next result is a classical inequality of Urysohn (for an elementary proof we
refer to [P]).
Lemma 2.2. Let B be a convex symmetric body in Rn . Then
Voln (B)
Voln (Bn2 )
n1
Z
≤
S n−1
||x||2B ∗
12
dλ(x)
ON THE VOLUME OF THE DOUBLE STOCHASTIC MATRICES
191
where Bn2 is the unit ball of `2n , B ∗ is the polar of B and λ is the normalized
Lebesgue measure on S n−1 .
It is well known that the latter integral can be expressed as a gaussian integral,
i.e.
Z
12
Z
12
1
||x||2B ∗ dλ(x)
=√
||x||2B ∗ dγn (x)
n
S n−1
Rn
where γn is the canonical gaussian probability measure on Rn .
Lemma 2.3. Let g1 , . . . , gk be not necessarily independent gaussian variables
with mean zero. Then
! 12
2
E sup |gj |
j≤k
1
≤ c(1 + log k) 2 sup ||gj ||2
j≤k
For a proof we refer to [P].
Lemmas 2.2 and 2.3 immediately imply the following
Proposition 2.4. Let x1 , . . . , xk be unit vectors in Rn . Then
Voln (co(x1 , . . . , xk )
Voln (Bn2 )
n1
r
≤c
log k
n
Proof. By Lemma 2.2 we have
Voln (B)
Voln (Bn2 )
n1

X
 12
1
≤ √ E||
gj ej ||2B ∗ 
n
j=1
where B is the absolutely convex hull of {x1 , . . . , xn } and (gj )nj=1 are independent
standard gaussian variables. Since
||x||B ∗ = sup |hxi , xi|
i≤k
we get from Lemma 2.3
E||
X
j=1
gj ej ||2B ∗ = E sup |
i≤k
X
gj hxi , ej i|2
j=1
≤ c(1 + log k)
192
M. SCHMUCKENSCHLÄGER
Theorem 2.5. There exists an absolute constant c such that the following inequalities hold.
1
2
cp
≤ Vol(n−1)2 (Dn ) (n−1)2 ≤
log n
n
n
Proof. Let P0 be the n × n matrix with the constant entry n1 in each place.
Then we conclude by Lemma 2.1
2
Vol(n−1)2 (Dn ) = Vol(n−1)2 [0, 1]n ∩ (E + P0 )
!
n2
1
1
= Vol(n−1)2
− ,1 −
∩E
n
n
!
n2
1 1
≥ Vol(n−1)2
− ,
∩E
n n
(n−1)2
2
≥
n
Thus the left hand side inequality is established. As for the right hand side observe
that for any permutation matrix Π:
Π − P0 ||HS =
2 ! 12
√
1
1
(n − n) 2 + n 1 −
= n−1
n
n
2
Since the number of permutation matrices is n! we deduce from the above proposition
s
! 1 2
(n−1)
p
√
Vol(n−1)2 (Dn )
log n!
≤ n − 1c
≤ c1 log n
2
2
2
Vol(n−1) (B(n−1)2 )
(n − 1)
Hence
1
(V ol Dn ) (n−1)2 ≤
c2 p
log n
n
References
[M-P] Meyer M. and Pajor A., Section of the unit ball of `pn , Journal of functional analysis 80
(1988).
[P]
Pisier G., The volume of convex bodies and Banach space geometry, Cambridge, 1989.
[S]
Schmuckenschläger M., Kahane type inequalities for subgroups of the orthogonal group,
preprint.
Vaaler D., A geometric inequality with applications to linear forms, Pacific J. Math. 83
[V]
(1979).
M. Schmuckenschläger, Johannes Kepler Universität, 4040 Linz, Austria
Download